how is the model exposed to the agents? is it a compact schema or raw text? we run ~10k concurrent agents and found that context window costs are the main bottleneck. feeding a full infrastructure graph into the context gets expensive fast.